Re-initialize emulator when refitting in AL

[radka-j]

This should follow the suggestion by @sgreenbury in #752

Similar approach is to be implemented in history matching workflow as part of #632 (see #748).

[sgreenbury]

For the AL tests, the learning rate was reduced in c55b4e8 to stability (not positive definite covariance matrices) when fitting the GP. It might also be good to explore how to set this hyperparameter in this issue for the AL workflow.